具有微电网的多样化、动态数据中心.pdf

1、Andrew A ChienUniversity of Chicago and Argonne National LabDiverse,Dynamic Datacenters with MicrogridsFuture Datacenters:Dynamic,Diverse with MicroGridsResearch Tools for Dynamic Datacenters Technologies for Thermal MicrogridsTechnologies for Electrical MicrogridsOutlineDataC Types:1-Many:AI,Cloud,

2、Equipment,DensityDynamic for Efficiency and Grid FlexibilityPower DistributionEfficient,Balanced,Oversubscribed=mGrid(complex mgmt)Heat RejectionEfficient,Balanced=mGrid(complex mgmt.)Diverse,Dynamic Datacenters with MicroGridsPowerHeatThermal uGridPower uGridDatacenterTypesAI TrainingMixed Train+In

3、fAI InferenceCloudNodeTypesGPUCPU-GPUCPUStorageDatacenter Use drives datacenter node types and mixMix varies from power-intensive nodes to less intensive nodesSpectrum continues to growDCGen 1.0:Generating Canonical IT ConfigurationsGenerative Model based on reference vendor system,datacenter,power,

4、cooling designsGenerates IT configurations based on target power,sqft,or IT hardware countGenerates power system and cooling system spec based on IT needs=Current and Projections to 2027,2030=In use today,public release Autumn 2025DCGen 1.0:Generating Canonical IT ConfigurationsIT(Nvidia,AMD,Dell,Su

5、permicro)Power,Cooling(Vertiv,Emerson,Schneider,Carrier,Johnson,)ReferenceDCs(xAI,Meta,)OCPdataDCGen1.0ITDesignPowerDesignCoolingDesignTarget:Power,sqft,IT HWConfig(JSON),SpaceReq,Power,etc.Gnibga&Chien,Datacenter Canonical IT Hardware Configurations,UChicago CS TR,July 2025.Objective:Explorepower d

6、ensities across DC types(current and future)Methodology:Fixed#of racks(10,000)Todays data centers:AI training DC at 79.8 kW/m2Mixed AI training and inference DC 53.5 kW/m2(1.5x lower)AI inference DC 18.8 kW/m2(4.2x lower)Cloud DC 10.4 kW/m2(7.7x lower)2027 data centers:power density increases signif